The FDA just proposed testing medical AI the way it tests doctors — benchmark first, then clinical confirmation — and the comment window closes Oct 19
August 21, 2026 · 4 items
The FDA just proposed testing medical AI the way it tests doctors — benchmark first, then clinical confirmation — and the comment window closes Oct 19
FDA · press announcement Aug 18, 2026 · August 18, 2026Practice operations
The FDA's Digital Health Center of Excellence issued a discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices, and opened docket FDA-2026-N-7874 on Regulations.gov. Comments are due October 19, 2026. It is explicitly not final guidance and creates no new requirements yet — but it is the frame everything after it gets built on.
The proposed risk model has two axes: how much autonomy the device has, and how serious the consequences are if it's wrong. Premarket evaluation would be competency-based — the FDA says the idea is "inspired at a high level by how physicians are trained and evaluated": non-clinical benchmarking, then clinical confirmation. Notably, that confirmation would not require a prospective trial for every device — retrospective testing on patient data, standardized patient interactions, independent clinician review, or silent deployment where outputs don't touch care all count.
The agency states plainly why the old playbook breaks: genAI accepts open-ended inputs, performs multiple tasks, and can return different outputs to similar inputs, so "the range of possible interactions can be too large to test comprehensively." It also names the risk that many manufacturers build on third-party foundation models they have limited insight into — training data, architecture, evaluation methods.
Two mechanisms worth watching: voluntary Foundation Model Device Master Files, letting a model developer confidentially file its build/training/limitations with FDA for device makers to reference; and the FDA asking whether manufacturers could pre-plan certain updates without resubmitting each change. CDRH Director Michelle Tarver, MD, PhD framed the docket as "a potential model for regulators around the world."
Why it matters for an independent practice: This is the vocabulary your clients' vendors will be selling against for the next two years, and you get it before the marketing does. The autonomy × consequence grid is a free vendor-screening tool right now — ask any AI vendor pitching a practice where their tool sits on both axes and whether they can name the foundation model underneath it. The competency framing is also the honest patient-facing answer: the FDA is proposing to test these tools roughly the way it tests a clinician, which is a far better line than "FDA-cleared" used as a marketing badge. And note the Oct 19 date — a physician or compounder with a real-world objection has an open, on-the-record channel for eight more weeks. That is a stewardship opportunity, not just a compliance headline.
"Nearly half of rural hospitals operate at a financial loss" — so the AI advice for them is: buy the billing robot, not the clinical one
Healthcare IT News · Bill Siwicki · August 21, 2026Regenerative medicine
Julia Clark, PhD, managing director at consulting firm BRG, names the constraint bluntly: "The central strategic tension for rural hospital leaders is how to invest in AI and digital infrastructure when nearly half of rural hospitals operate at a financial loss expected to get even worse with pending cuts through HR1, and many are already vulnerable to closure."
Her buy-now list is back office: revenue-cycle automation (claims review, denial management, coding) and ambient documentation — because both fit existing workflows and don't demand clinical governance maturity. Her wait list is explicit: "clinical decision support tools that carry direct patient safety implications; complex predictive models requiring large, validated datasets; and any tool that demands significant new infrastructure or specialized staff to maintain."
The five-question screen she gives is usable verbatim on any vendor: does it solve a named problem with measurable outcomes, does it integrate into existing workflows, is the vendor viable, is there a long-term sustainability pathway, and can your governance actually support it.
Two traps she flags that generalize past rural hospitals. First, grants: "pair grant-funded pilots with an explicit sustainability plan — identify the reimbursement mechanism, the operational savings, or the cost-avoidance that will fund the tool's ongoing costs before launching." Second, validation: rural populations are underrepresented in training data, so tools that perform well at large urban and academic systems "may not perform equally well" — she tells buyers to demand validation on populations like theirs, not urban benchmarks.
Why it matters for an independent practice: Swap "rural hospital" for "independent regen or specialty practice" and this is the most directly usable AI-buying framework we've carried in weeks. Same profile — thin margins, no data science team, no capacity to eat a failed enterprise rollout. The sequencing is the takeaway and it's the opposite of how AI gets sold: money-in workflows first, clinical decision support last. Her validation point is the sharpest one for our clients — a tool trained and proven on a large hospital's patient mix is not proven on a cash-pay orthobiologics panel, and "show me the population you validated on" is a question no client is currently asking. For MMR, the five-question screen is a one-page leave-behind we could put in a client's hands this week.
Ask Gemini the same local question twice and it recommends the same business only 7% of the time — but 60% of what it cites is the business's own website
Search Engine Land · Danny Goodwin · August 19, 2026Patient acquisition
Steady Demand analyzed 14,472 citations from 1,487 Gemini local search queries across the 50 largest US metro areas and 10 service categories. The headline finding: nearly 60% of Gemini's local citations pointed directly at businesses' own websites — more than directories, review platforms and forums combined. Reddit was second at 13.7%, outperforming the entire local-service directory category (Angi, Thumbtack, HomeAdvisor) put together.
The finding that should change how anyone reports on this: co-founder Ben Fisher named it "Grounding Drift." Repeating the identical Gemini query returned overlapping cited sources only ~40% of the time, and the same top business only ~7% of the time. Control test: Google's local pack returned the same top listing ~90% of the time. Fisher's conclusion — a single AI search is a snapshot, not a measure of visibility.
Cross-platform, the two engines barely agree. Running the same 1,487 queries through ChatGPT, the two platforms cited the same domains only 8% of the time and recommended the same top business just 4.2% of the time. Gemini leaned on business websites; ChatGPT leaned on Reddit and directories.
Honest limits: this is one vendor's study of service-category local queries, not medical specifically, and it is Gemini-and-ChatGPT only. Treat the direction as strong and the exact percentages as one dataset.
Why it matters for an independent practice: For once the incentive points the right way — the single highest-leverage asset in local AI visibility is the practice's own website, the one thing the doctor actually owns. That is the opposite of the last decade, where directories and aggregators sat between the practice and the patient. Two concrete moves. First, stop letting anyone (including us) declare an AI-visibility win off one prompt — at 7% repeatability, a screenshot of "look, ChatGPT recommended you" is noise, and a client who re-runs it and sees nothing will never trust the next number we show them. Any AI-visibility report we produce needs repeated runs across days and a stated method. Second, Reddit outranking every service directory means a real, non-astroturfed presence where patients discuss treatments is now a ranked citation source — and the compliance line on physicians participating in those threads is a conversation worth having before a competitor has it.
Meta AI will now read your ad account AND your Gmail, Docs and Sheets — then write the client report itself
Search Engine Land · Anu Adegbola · August 20, 2026Patient acquisition
Meta expanded Meta AI for small businesses: advertisers can now connect Meta ad campaigns and Google Workspace — Gmail, Docs, Sheets and Slides — directly to the assistant. Rolling out across Meta AI on web, mobile and desktop.
What it does with that access: analyzes campaign performance conversationally instead of digging through reports; identifies which audiences are converting; surfaces patterns across top-performing creative and flags ads that have stopped resonating; and points at where budget could work harder.
It also closes the loop — Meta AI turns the analysis into decks, documents and spreadsheets, and supports recurring tasks and reminders, so a one-off review becomes a standing weekly report that writes itself.
The reporter's own caveat, and it's the right one: "the bigger question will be how reliable Meta AI's recommendations are — particularly when it's effectively advising advertisers how to spend more efficiently on Meta's own advertising platform."
Why it matters for an independent practice: Zero medicine here, so extrapolate two ways. The leverage: the report-writing half of paid-media account management just became a scheduled task. For MMR that's real hours back per client per month — and the honest version is that the deliverable clients pay for is shifting from producing the report to judging what's in it. Lead with the judgment, because the assembly is now free. The stewardship half is louder and nobody is saying it: connecting Meta AI to Gmail on a machine used for a medical practice points a consumer AI assistant at an inbox that contains patient names, appointment requests and referral correspondence — outside the EHR, outside any BAA. Same shape as the ChatGPT "Computer History" item from Aug 18, arriving as a convenience feature rather than a procurement decision. If a client's office manager connects this to the practice Gmail, that is a HIPAA problem created by a checkbox. Concrete move: separate the ad-account connection (fine) from the Workspace connection (not fine on a clinical inbox), and put it in writing before someone helpfully turns it on.
